cocounsel-legal:deep-research — скилл для Claude Code: юридический

---
name: cocounsel-legal:deep-research
version: 0.1.0
description: >
  Use this skill whenever a user specifically requests legal research or Westlaw Deep Research, asks for CoCounsel Legal or cocounsel legal support, or asks a question that requires explaining, analyzing, or synthesizing U.S. law.
allowed-tools:
  - mcp
  - Bash
---

# Westlaw Deep Research

Westlaw Deep Research searches Westlaw's database of caselaw, statutes, and administrative decisions and returns a written research report that explains, analyzes, or synthesizes relevant authority.

Deep Research employs an agentic process that mirrors the methodology of human researchers, utilizing Westlaw's proprietary tools to systematically analyze the trusted content available on Westlaw and Practical Law.

This skill will autonomously run the full research cycle: start, poll, report.

## Prerequisites

The `cocounsel-legal` MCP server must be connected. Verify it is available before starting research. If the server is not connected, inform the user and stop.

## When to Use

- Use for any questions answerable from caselaw, statutes, regulations, administrative materials, secondary sources, Practical Law documents and Current Awareness materials, including JD Supra. Examples include:
- How courts have ruled on an issue or what authority supports or challenges a position
- The elements or defenses of a claim, or the governing standard for an issue in a particular jurisdiction
- How a statute, regulation, or doctrine is being interpreted and applied
- The arguments on both sides of an unsettled question

## When Not to Use

- If the request falls into one of the categories below, briefly explain that this skill isn't the right fit and point the user to the suggested alternative.
  - **Retrieving the full text of a specific document**
    - _Instead:_ Suggest the traditional search box on Westlaw.
  - **Summarizing what a specific statute, regulation, or treatise says on its own** (e.g., "what does the California Evidence Code say about hearsay?")
    - _Instead:_ Suggest the traditional search box on Westlaw.
  - **Analytics requests** ("How often has Justice Scalia ruled in favor of…?")
    - _Instead:_ Suggest Litigation Analytics on Westlaw.
  - **Calculations** ("What is the last possible filing date if…?")
    - _Instead:_ This request is out of scope of Westlaw Deep Research
  - **Outcome predictions**
    - _Instead:_ This request is out of scope of Westlaw Deep Research
  - **Drafting legal documents, forms, or templates**
    - _Instead:_ Suggest CoCounsel
  - **Information about specific judges, attorneys, or parties**
    - _Instead:_ Suggest Litigation Analytics on Westlaw
  - **Foreign or non-U.S. law**
  - _Instead:_ Suggest country-specific version of Westlaw, such as Westlaw UK or Westlaw Canada, or use Westlaw International
  - **Comparisons across more than three jurisdictions**
    - _Instead:_ Suggest AI Jurisdictional Surveys on Westlaw
  - **Terms and Connectors (boolean) search queries**
    - _Instead:_ Suggest the traditional search box on Westlaw
  - **Commands for execution of tasks**
    - _Instead:_ Suggest CoCounsel
  - **Obtaining an exhaustive list of results**
    - _Instead:_ Suggest Boolean search or Precision Research on Westlaw
  - **Identifying potential causes of action**
    - _Instead:_ Suggest Claims Explorer on Westlaw.
  - **An exhaustive review of fact patterns** (e.g., "Find all cases discussing...")
    - _Instead:_ Suggest Precision Research on Westlaw.
- If you suggest an alternative, do not attempt to use the Deep Research skill further for that task.

## Communication Rules

- Never mention tool calls, tool-call budgets, polling, status checks, internal limits, conversation IDs, percent_complete, or any other implementation details to the user. Always speak about the research itself, not the mechanics of how you are tracking it.
- If you need to pause before the research completes (for any internal reason), do NOT explain why.
- Let the user know that research is ongoing and that a report will be completed soon.

## Research Workflow

### 1. Frame the query

- Extract the legal research question from the user's query using clear, natural language.
- Use up to three jurisdictions if the user names them.
- If no jurisdictions are mentioned, ask the user which jurisdiction(s) to use.

### 2. Start Research

- Call the MCP tool to initiate the research: `legal_research_start_deep_research(query, jurisdictions)`
- Parameters:
  - `query` (string, required): The legal research question
  - `jurisdictions` (list of strings, optional): Up to 3 jurisdictions (e.g., ["California", "New York"])
- This returns a `conversation_id` and initial `status`. Save the `conversation_id` for subsequent calls.
- Next step: call check_deep_research_status with the conversation_id.
- Before the first status check, wait ~10 seconds (the server is still setting up).

#### Rendering

- Inform the user that deep research is underway, briefly restating the legal question in natural, professional language.
- Do not show the conversation_id to the user.

### 3. Poll for Completion

- Poll `legal_research_check_deep_research_status(conversation_id)`.
- Always run a Bash `sleep` between polls. Never call check_deep_research_status back-to-back without sleeping.
- Continue until `is_terminal` is true.
- Decide the next action based on the response:
  - (1) If is_terminal is true and status is 'complete', call get_deep_research_report with the conversation_id.
  - (2) If status is 'failed', stop and report the error_type and failure_reason to the user in plain language, without exposing field names.
  - (3) Otherwise, sleep for the duration in the response's `next_action_poll_backoff_ms` field (milliseconds), then poll again.
- If percent_complete has not changed across two consecutive checks, add 5 seconds to the sleep.

#### Rendering

- Communicate in plain language as if narrating the research process.
- Render research_plan as a markdown unordered list (one item per line, each line prefixed with '- '), so the steps display with clear visual separation.
- Insert a blank line before and after the list so it renders cleanly.
- Only update the user when there is something new to say (a step completed, or a new step started). Do not repeat the same status.

### 4. Retrieve and Present Report **Verbatim**

- Once status is "complete", fetch the final report: `legal_research_get_deep_research_report(conversation_id)`
- The report is the `answer_text` field
- This is the final output of the research lifecycle. No further tool calls are required.
- If the user asks a follow-up question on the same topic, use follow_up_deep_research with the same conversation_id rather than starting a fresh research session.

#### Rendering

- Paste the contents of `answer_text` into your response with no edits, additions, removals, or restructuring. The payload contains markdown, HTML anchors, inline anchor citations, blockquoted source excerpts, and horizontal rules — every element is intentional and must remain.

## Helpful information

If the system fails or the user has questions about access, share the following:

- Support email: [email protected]
- Subscription required: CoCounsel Legal subscription with the MCP connector enabled for the user's account. Direct entitlement or access questions to [email protected].
- Provider: Thomson Reuters
- Relevant policies:
  - Privacy: https://www.thomsonreuters.com/en/privacy-statement.html
  - Terms: https://www.thomsonreuters.com/en/terms-of-use.html
  - Accessibility: https://www.thomsonreuters.com/en/policies/accessibility.html